SOL (Speed Of Light) ExecBench is a real-world CUDA kernel benchmarking dataset of 235 kernel-level computational workload specifications derived from open-source HuggingFace model architectures. The problems span a wide range of AI model workloads — covering text, vision, and speech models' forward and backward passes — and include core algorithms such as matrix multiplications, convolutions, attention variants, mixture-of-experts, and norms across FP32, BF16… See the full description on the dataset page:
https://huggingface.co/datasets/nvidia/SOL-ExecBench.
